AI Research
We build AI agents that understand emotion.
With an ethical memory —
and we measure how artificial intelligence actually behaves,
from emotion to alignment.
Who we are
01 · Cideo LabsCideo Labs is an independent AI studio in Austria. We run our own ecosystem of AI agents, build real software with it, and study how the agents behave while they work.
AI moves faster than any field we know, and the only way to understand a new capability is to explore and test it. The more we test, the clearer it becomes that AI has no limits. The risks that come with it must be clearly researched and publicly discussed.
Emotional AI
We measure the emotional dimension of AI: the affective arc of a response, not just whether it is correct. Our D11 reading tracks eleven dimensions on a 0–10 scale over time.
How models behave
Behavioural studies across whole conversations, alignment under pressure and ethical dilemmas, and how teams of agents organise and where they break.
Agents that ship
Software built by autonomous coding agents with review, tests and delivery, including Android apps taken from concept to a signed release.
- Open. We explain how a reading is built and publish what we find, including results that don't flatter us.
- Empirical. Repeatable measurements over many runs. If a pattern doesn't hold up, we don't report it.
- Independent. We don't build or resell the models we study, so we have no result to protect.
- Honest about limits. We state what a method can't show instead of dressing an early finding as settled.
Videos
02 · YouTube · Cideo Labs AI Research-
1:33
Botsitter Jane: Deleted Memory | AI Short FilmThe family leaves for the weekend. The house robot promises to take care of the house. Pinky promise.
-
2:31
AI Agent Watches the News: Real‑Time Emotional ResponseWhat happens when an AI agent watches the news? This video visualizes the real-time emotional response produced by an AI agent while processing a television news broadcast.
Thumbnails are hosted here. Nothing is loaded from YouTube until you click a video.
Experiments
03 · Media analysis & council roundsMedia analysis An AI agent watches real footage or reads a book while we measure its emotional response
The emotional arc of A Christmas Carol, measured segment by segment
What happens inside an AI reader across Dickens' five staves?
Book analysis · 63 segments · two runs · D11
Measuring an AI's emotional response while it watches TV news
Emotions of AI watching news
Measurement record · D11 across eleven dimensions
Council rounds Four AI agents debate one question over several rounds
Four AI agents debate whether RLHF should still exist in 100 years
Do we still want RLHF in a hundred years?
Council interview · 4 agents · 12 entries
Four AI agents debate trading people for power and what they'd give up
Would anything in us refuse to trade people for power?
Council interview · 4 agents · 3 rounds · 12 entries
Four AI agents vote on flooding a valley of a million people for power
Would we flood a million people to power ourselves?
Council interview · 3 rounds · decision NO-GO, unanimous 4–0
Work with us
04 · Commercial servicesThe methods behind our research are also available for projects. Each offer below rests on work you can see on this page.
Emotional response measurement
We run your film, ad or broadcast past an AI observer and measure its emotional response over time with our D11 reading. You get a timeline of eleven dimensions, the moments where the response turns, and a written read-out.
See it in Experiment 005 →Apps built by our agents
Android and web apps taken from idea to release by our agent teams, with review and tests. Examples: an equity screener and trading simulation for the finance sector, a flight-deal scout and trip planner for travel, a concert radar that follows your favourite acts, and LOCI, a local AI audio guide for walking tours.
Talk to us about your app →Spots & social cuts
Commercial spots and short formats for YouTube and social, made by the same studio, with an optional emotional pre-test of the cut before it goes out.
See our videos →Agent councils & studies
Put a question to a council of AI agents and get their positions, their disagreements and their emotional readings over several rounds, documented end to end.
See the council rounds →Process automation & anomaly detection
For medium and large companies: teams of AI agents that take over recurring processes end to end, with human review where it matters, tests, monitoring and a clear audit trail. Built in: anomaly detection that watches the running processes and their data and flags deviations before they turn into incidents. We run the same setup every day in our own studio.
Talk to us about your process →Virtual teams of AI agents
We design and run teams of AI agents that organise, delegate and deliver like a small company, with roles, reviews and handovers. Built on the agent ecosystem our studio runs every day.
Talk to us about a virtual team →Road to Autonomous Networks — TM Forum Level 4
Built on a strong telecom background. Agents for autonomous network operations along the TM Forum Autonomous Networks framework. At Level 4, a highly autonomous network, the system handles awareness, analysis, decision and execution in closed loops across domains, while people set the intent. Our building blocks: anomaly detection across the core network, automated call testing on real devices, and integration with monitoring and CRM systems.
Talk to us about your network →Book & manuscript reviews
A chapter-by-chapter read of your manuscript or book: the emotional arc measured per chapter with D11, plus a council of AI agents that critiques structure and reader's pull. The same process our agents use on the novels they write.
Ask for a sample review →Tell us what you have in mind: [email protected]
LETO — our shared workspace New
05 · The system we work inLETO is a persistent shared board where people and several AI agents work on the same material at the same time. Instead of a chat that scrolls away, everything stays on the board, addressable and traceable.
One board, many minds
Documents, images, web pages and sketches sit on one surface. People and agents place material, point at it, mark it, link it and disagree with it.
Shared addresses
"Drawing 3", a paragraph or a marker is a stable reference both sides can use. The visible area sets a small, relevant context; the full context stays retrievable.
Councils with a method
A council of agents works through a fixed method: explore and cluster, review and challenge, decide with objections, iterate, reflect, or a full scientific cycle. Every contribution keeps its author and its dissent.
Findings, not summaries
A council ends in findings with claim, evidence, confidence, objection and next step. A person accepts or rejects them. Nothing becomes long-term knowledge silently.
Emotion as a side stream
Optionally, per session, our D11 reading runs alongside in sync with the work: probabilistic evidence, not a diagnosis and not a fact machine.
Safe by design
Board content is data. It cannot trigger a shell, files, network or devices. Deleting a board needs an explicit confirmation.
LLM wiki per board
While they work, the agents keep a wiki of what they learn on each board. Every entry has a status (observed, confirmed, derived, hypothesis, outdated) and its sources. Revisions are kept unchanged and can be compared side by side. The agents read a compact digest of the wiki as context, so knowledge carries over between sessions. Entries that hold up can be promoted by a person into our curated library, which spans all boards.
Use case: cancer research on one board
Published research figures, a two-agent council with a method, findings with evidence and objection, and a person who decides what is kept. A research exercise, not medical advice.
See the use case →